Abstract Reliable condition monitoring of permanent magnet synchronous motor (PMSM) encoders under extreme industrial noise remains a persistent challenge due to severe feature corruption. To address this, this paper proposes a Physics-Informed Multimodal Front-end Enhancement (PI-MFE) framework. The framework systematically internalizes physical priors as inductive biases into the feature extraction hierarchy to actively disentangle nonlinear noise distortions. For the electromagnetic modality, a Physics-Constrained Dual Attention Network (PCDAN) is orchestrated by a novel spectral consistency constraint. This mechanism explicitly enforces the latent manifold topology to align with authentic phase-modulation harmonics, fundamentally breaking the inherent spectral bias of deep networks. Concurrently, for the mechanical vibration modality, a Hybrid PSO Physics-Informed Neural Network (HPSO-PINN) is formulated as an uncertainty-aware probabilistic surrogate. By adaptively allocating computational queries driven by cognitive uncertainty, this strategy elegantly reformulates the intractable non-convex fitness landscape into an active learning paradigm, decisively circumventing the combinatorial explosion of heuristic evaluations during mode decomposition. A minimalist Linear SVM decodes the enhanced features, with the proposed PI-MFE(Linear-SVM) attaining peak accuracies of 98.1 ± 0.1% in simulation and 95.8 ± 0.3% on the physical platform, while retaining 84.8 ± 0.2% and 82.3 ± 0.5% fusion accuracy under the extreme -5 dB condition, respectively. The small accuracy gap Δ between Linear SVM and RBF-SVM further indicates effective linear separability, confirming that physics-driven front-end enhancement enables robust edge diagnostics.
原文链接
分享
引用
